Multi-Length CSI Feedback With Ordered Finite Scalar Quantization

We propose a novel, lightweight, deep-learning based model, which enables fast, multi-length channel state information (CSI) feedback. The proposed method harnesses the advantages of finite scalar quantization and ordered representation learning, to create the ordered finite scalar quantization (OFSQ) scheme, which has a simple structure, with significantly reduced complexity, while demonstrating solid CSI reconstruction ability for any desired feedback bitstream length. Our method reshapes latent vectors into sub-vectors, applies a hyperparameter-based and bounded scalar quantization, while it integrates a nested dropout layer to prioritize sub-vectors based on their importance to CSI retrieval. Simulation results confirm that the proposed scheme significantly reduces the computational complexity, as it avoids to exhaustively search the quantization codebook, while it shows an improved CSI reconstruction ability compared to state-of-the-art multi-length CSI feedback models. Therefore, OFSQ is a promising plug-in architecture, which can be paired with any autoencoder for use in wireless communication systems.